ChainerRL: A Deep Reinforcement Learning Library
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Journal of Machine Learning Research 22 (2021) 1-14 Submitted 4/20; Revised 12/20; Published 4/21
ChainerRL: A Deep Reinforcement Learning Library
Yasuhiro Fujita fujita@preferred.jp
Prabhat Nagarajan prabhat@preferred.jp
Toshiki Kataoka kataoka@preferred.jp
Preferred Networks
Tokyo, Japan
Takahiro Ishikawa sykwer@g.ecc.u-tokyo.ac.jp
The University of Tokyo
Tokyo, Japan
Editor: Andreas Mueller
Abstract
In this paper, we introduce ChainerRL, an open-source deep reinforcement learning (DRL)
library built using Python and the Chainer deep learning framework. ChainerRL implements
a comprehensive set of DRL algorithms and techniques drawn from state-of-the-art research
in the field. To foster reproducible research, and for instructional purposes, ChainerRL
provides scripts that closely replicate the original papers’ experimental settings and reproduce
published benchmark results for several algorithms. Lastly, ChainerRL offers a visualization
tool that enables the qualitative inspection of trained agents. The ChainerRL source code
can be found on GitHub: https://github.com/chainer/chainerrl.
Keywords: reinforcement learning, deep reinforcement learning, reproducibility, open
source software, chainer
1. Introduction
Since its resurgence in 2013 (Mnih et al., 2013), deep reinforcement learning (DRL) has
undergone tremendous progress, and has enabled significant advances in numerous complex
sequential decision-making problems (Mnih et al., 2015; Silver et al., 2018; Levine et al., 2016;
Kalashnikov et al., 2018). The machine learning community has witnessed a growing body of
literature on DRL algorithms (Henderson et al., 2018). However, coinciding with this rapid
growth has been a growing concern about the state of reproducibility in DRL (Henderson
et al., 2018). The growing body of algorithms and increased reproducibility concerns beget
the need for comprehensive libraries, tools, and implementations that can aid RL-based
research and development.
Many libraries aim to address these challenges in different ways. rllab (Duan et al., 2016)
and its successor, garage, provide systematic benchmarking of continuous-action algorithms
on their own benchmark environments. Dopamine (Castro et al., 2018) primarily focuses
on DQN and its extensions for discrete-action environments. rlpyt (Stooke and Abbeel,
2019) supports both discrete and continuous-action algorithms from the three classes: policy
gradient (with V-functions), deep Q-learning, and policy gradient with Q-functions. Other
libraries also support diverse sets of algorithms (Dhariwal et al., 2017; Caspi et al., 2017;
Hill et al., 2018; Liang et al., 2018). catalyst.RL (Kolesnikov and Hrinchuk, 2019) aims
c 2021 Yasuhiro Fujita, Prabhat Nagarajan, Toshiki Kataoka, and Takahiro Ishikawa.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided
at http://jmlr.org/papers/v22/20-376.html.
Fujita, Nagarajan, Kataoka, and Ishikawa
ActionValues
Discrete, Categorical, Explorers
Quantile, Quadratic ε-greedy, Boltzmann, Additive
ChainerRL Visualizer
Gaussian or Ornstein-Uhlenbeck
Distributions
Deterministic, Gaussian,
Softmax, Mellowmax Experiment Utilities
Serial training and synchronous or
Agents asynchronous parallel training for any OpenAI
Predefined Arch Gym-like environment
Neural Networks DQN, C51, Rainbow, IQN
DQN, Dueling, MLP
A2C, A3C, PPO, SAC
DDPG, TD3, ACER, PCL, etc.
Recurrent Networks
Paper Reproductions
Replay Buffers Full reproduction scripts and results for
Noisy Networks Support prioritized, N-step, and/or episodic ChainerRL A3C, DQN, IQN, Rainbow, DDPG, PPO,
sampling TRPO, TD3, and SAC
Figure 1: A depiction of ChainerRL. Using ChainerRL’s building blocks, DRL algorithms,
called agents, are written by implementing the Agent interface. Agents can
be trained with the experiment utilities and inspected with the ChainerRL
Visualizer.
to address reproducibility issues in RL via deterministic evaluations and by tracking code
changes for continuous-action algorithms.
In this paper, we introduce ChainerRL, an open-source Python DRL library supporting
both CPU and GPU training, built off of the Chainer (Tokui et al., 2019) deep learning
framework. ChainerRL offers a comprehensive set of algorithms and abstractions, a set of
“reproducibility scripts” that replicate research papers, and a companion visualizer to inspect
agents.
2. Design of ChainerRL
In this section, we describe ChainerRL’s design, as in Figure 1.
2.1 Agents
In ChainerRL, each DRL algorithm is written as a class that implements the Agent interface.
The Agent interface provides a mechanism through which an agent interacts with an en-
vironment, e.g., through an abstract method Agent.act_and_train(obs, reward, done)
that takes as input the current observation, the previous step’s immediate reward, and a flag
for episode termination, and returns the agent’s action to execute in the environment. By
implementing such methods, both the update rule and the action-selection procedure are
specified for an algorithm.
An agent’s internals consist of any model parameters needed for decision-making and
model updating. ChainerRL includes several built-in agents that implement key algorithms
including the DQN (Mnih et al., 2015) family of algorithms, as well as several policy gradient
and actor-critic algorithms.1
1. ChainerRL’s algorithms include: DQN (Mnih et al., 2015), Double DQN (Van Hasselt et al., 2016),
Categorical DQN (Bellemare et al., 2017), Rainbow (Hessel et al., 2017), Implicit Quantile Networks
(IQN) (Dabney et al., 2018), Off-policy SARSA, (Persistent) Advantage Learning (Bellemare et al.,
2ChainerRL: A Deep Reinforcement Learning Library
2.2 Experiments
While users can directly interact with agents, ChainerRL provides an experiments module
that manages agent-environment interactions as well as training/evaluation schedules. This
module supports any environment that is compatible with OpenAI Gym’s Env (Brockman
et al., 2016). An experiment takes as input an agent and an environment, queries the agent
for actions, executes them in the environment, and feeds the agent the rewards for training
updates. Moreover, an experiment can periodically perform evaluations and collect evaluation
statistics. Through the experiments module, ChainerRL supports batch or asynchronous
training, enabling agents to act, train, and evaluate synchronously or asynchronously in
several environments in parallel. A full list of synchronous and asynchronous agents is
provided in the appendix.
2.3 Developing a New Agent
The Agent interface is defined very abstractly and flexibly so that users can easily implement
new algorithms while leveraging the experiments utility and parallel training infrastructure.
To develop a new agent, we first create a class that inherits Agent. Next, the learning
update rules and the agent’s action-selection mechanisms are implemented using ChainerRL’s
provided building blocks (see Section 2.4). Once an agent is created, the agent and a Gym-like
environment can be given to the experiments module to easily train and evaluate the agent
within the specified environment.
2.4 Agent Building Blocks
ChainerRL offers a set of reusable components for building new agents, including ChainerRL’s
built-in agents. Though not comprehensive, we highlight here some of the building blocks
that demonstrate the flexibility and reusability of ChainerRL.
Explorers For building action-selection mechanisms during training, ChainerRL has built-
in explorers including -greedy, Boltzmann exploration, additive Gaussian noise, and
additive Ornstein-Uhlenbeck noise (Lillicrap et al., 2016).
Replay buffers Replay buffers (Lin, 1992; Mnih et al., 2015) have become standard tools
in off-policy DRL. ChainerRL supports traditional uniform-sampling replay buffers,
episodic buffers for sampling past (sub-)episodes for recurrent models, and prioritized
buffers that prioritize sampled transitions (Schaul et al., 2016). ChainerRL also supports
sampling N steps of transitions, for algorithms based on N -step returns.
Neural networks While ChainerRL supports any Chainer model, it has several pre-defined
architectures, including DQN architectures, dueling network architectures (Wang et al.,
2016), noisy networks (Fortunato et al., 2018), and multi-layer perceptrons. Recurrent
models are supported for many algorithms, including DQN and IQN.
2016), (Asynchronous) Advantage Actor-Critic (A2C (Wu et al., 2017), A3C (Mnih et al., 2016)),
Actor-Critic with Experience Replay (ACER) (Wang et al., 2017), Deep Deterministic Policy Gradients
(DDPG) (Lillicrap et al., 2016), Twin-delayed double DDPG (TD3) (Fujimoto et al., 2018), Proximal
Policy Optimization (PPO) (Schulman et al., 2017), REINFORCE (Williams, 1992), Trust Region Policy
Optimization (TRPO) (Schulman et al., 2015), and Soft Actor-Critic (SAC) (Haarnoja et al., 2018).
3Fujita, Nagarajan, Kataoka, and Ishikawa
Distributions Distributions are parameterized objects for modeling action distributions.
Network models that return Distribution objects are considered policies. Supported
policies include Gaussian, Softmax, Mellowmax (Asadi and Littman, 2017), and deter-
ministic policies.
Action values Similar to Distributions, ActionValues parameterizing the values of ac-
tions are used as outputs of neural networks to model Q-functions. Supported Q-
functions include the standard discrete-action Q-function typical of DQN as well as
categorical (Bellemare et al., 2017) and quantile (Dabney et al., 2018) Q-functions
for distributional RL. For continuous action spaces, quadratic Q-functions called
Normalized Advantage Functions (NAFs) (Gu et al., 2016) are also supported.
By combining these agent building blocks, users can easily construct complex agents such
as Rainbow (Hessel et al., 2017), which combines six features into a single agent. This ability
is highlighted in Appendix D, which provides a pseudocode construction of a Rainbow agent
and trains it in multiple parallel environments in just a few lines.
2.5 Visualization
ChainerRL is accompanied by the ChainerRL Visualizer, which takes as input an environ-
ment and an agent, and enables users to easily inspect agents from a browser UI. With the
visualizer, one can visualize the portions of the pixel input that the agent is attending to
as a saliency map (Greydanus et al., 2018). Additionally, users can either manually step
through the episode or view full rollouts of agents. Moreover, the visualizer depicts the
probabilities with which the agent will perform specific actions. If the agent learns Q-values
or a distribution of Q-values, the predicted Q-value or Q-value distribution for each action
can be displayed. Figure 2 in Appendix C depicts some of these features.
3. Reproducibility
Many DRL libraries offer implementations of algorithms but often deviate from the original
paper’s implementation details. We provide a set of “reproducibility scripts”, which are
compact examples (i.e., single files) of paper implementations written with ChainerRL that
match, as closely as possible, the original paper’s (or in some cases, another published
paper’s) implementation and evaluation details. ChainerRL currently has “reproducibility
scripts” for DQN, IQN, Rainbow, A3C, DDPG, TRPO, PPO, TD3, and SAC. For each of
these algorithms and domains, we have released pretrained models for every domain, totaling
hundreds of models. Moreover, for each script, we provide full tables of our scores and
compare them against scores reported in the literature (Tables 2 and 4 in Appendix B).
4. Conclusion
This paper introduced ChainerRL and the ChainerRL Visualizer. ChainerRL’s comprehen-
sive suite of algorithms, flexible APIs, visualization tools, and faithful reproductions can
accelerate the research and application of DRL algorithms. While ChainerRL targets Chainer
users, we have developed an analogous library, PFRL, for PyTorch users.2
2. The PFRL code is located at https://github.com/pfnet/pfrl.
4ChainerRL: A Deep Reinforcement Learning Library
Acknowledgments
We thank Avinash Ummadisingu, Mario Ynocente Castro, Keisuke Nakata, Lester James V.
Miranda, and all the open source contributors for their contributions to the development of
ChainerRL. We thank Kohei Hayashi and Jason Naradowsky for useful comments on how to
improve the paper. We thank the many authors who fielded our questions when reproducing
their papers, especially George Ostrovski.
References
Kavosh Asadi and Michael L. Littman. An Alternative Softmax Operator for Reinforcement
Learning. In ICML, 2017.
Marc G Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The Arcade Learning
Environment: An Evaluation Platform for General Agents. Journal of Artificial Intelligence
Research, 47:253–279, 2013.
Marc G. Bellemare, Georg Ostrovski, Arthur Guez, Philip S. Thomas, and Rémi Munos.
Increasing the Action Gap: New Operators for Reinforcement Learning. In AAAI, 2016.
Marc G. Bellemare, Will Dabney, and Rémi Munos. A Distributional Perspective on
Reinforcement Learning. In ICML, 2017.
Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie
Tang, and Wojciech Zaremba. OpenAI Gym. arXiv preprint arXiv:1606.01540, 2016.
Itai Caspi, Gal Leibovich, Gal Novik, and Shadi Endrawis. Reinforcement learning coach,
December 2017. URL https://doi.org/10.5281/zenodo.1134899.
Pablo Samuel Castro, Subhodeep Moitra, Carles Gelada, Saurabh Kumar, and Marc G.
Bellemare. Dopamine: A Research Framework for Deep Reinforcement Learning. 2018.
Will Dabney, Georg Ostrovski, David Silver, and Rémi Munos. Implicit Quantile Networks
for Distributional Reinforcement Learning. In ICML, 2018.
Prafulla Dhariwal, Christopher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec
Radford, John Schulman, Szymon Sidor, Yuhuai Wu, and Peter Zhokhov. OpenAI Baselines.
https://github.com/openai/baselines, 2017.
Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking
Deep Reinforcement Learning for Continuous Control. In ICML, 2016.
Meire Fortunato, Mohammad Gheshlaghi Azar, Bilal Piot, Jacob Menick, Ian Osband, Alex
Graves, Vlad Mnih, Remi Munos, Demis Hassabis, Olivier Pietquin, Charles Blundell, and
Shane Legg. Noisy Networks for Exploration. In ICLR, 2018.
Scott Fujimoto, Herke van Hoof, and Dave Meger. Addressing Function Approximation Error
in Actor-Critic Methods. In ICML, 2018.
5Fujita, Nagarajan, Kataoka, and Ishikawa
Sam Greydanus, Anurag Koul, Jonathan Dodge, and Alan Fern. Visualizing and Under-
standing Atari Agents. In ICML, 2018.
Shixiang Gu, Timothy Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous Deep
Q-Learning with Model-based Acceleration. In ICML, 2016.
Tuomas Haarnoja, Henry Zhu, George Tucker, and Pieter Abbeel. Soft Actor-Critic Algo-
rithms and Applications. arXiv preprint arxiv:1812.05905, 2018.
Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David
Meger. Deep Reinforcement Learning that Matters. In AAAI, 2018.
Matteo Hessel, Joseph Modayil, Hado Van Hasselt, Tom Schaul, Georg Ostrovski, Will
Dabney, Dan Horgan, Bilal Piot, Mohammad Azar, and David Silver. Rainbow: Combining
Improvements in Deep Reinforcement Learning. In AAAI, 2017.
Ashley Hill, Antonin Raffin, Maximilian Ernestus, Rene Traore, Prafulla Dhariwal, Christo-
pher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec Radford, John Schul-
man, Szymon Sidor, and Yuhuai Wu. Stable baselines. https://github.com/hill-a/
stable-baselines, 2018.
Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang,
Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine.
QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation.
In CoRL, 2018.
Sergey Kolesnikov and Oleksii Hrinchuk. Catalyst.RL: A Distributed Framework for Repro-
ducible RL Research. arXiv preprint arXiv:1903.00027, 2019.
Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-End Training of
Deep Visuomotor Policies. The Journal of Machine Learning Research, 17(1):1334–1373,
2016.
Eric Liang, Richard Liaw, Robert Nishihara, Philipp Moritz, Roy Fox, Ken Goldberg,
Joseph Gonzalez, Michael Jordan, and Ion Stoica. RLlib: Abstractions for Distributed
Reinforcement Learning. In ICML, 2018.
Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval
Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement
learning. In ICLR, 2016.
Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and
teaching. Machine Learning, 8(3-4):293–321, 1992.
Marlos C Machado, Marc G Bellemare, Erik Talvitie, Joel Veness, Matthew Hausknecht,
and Michael Bowling. Revisiting the Arcade Learning Environment: Evaluation Protocols
and Open Problems for General Agents. Journal of Artificial Intelligence Research, 61:
523–562, 2018.
6ChainerRL: A Deep Reinforcement Learning Library
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan
Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. In NIPS
Deep Learning Workshop, 2013.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei a Rusu, Joel Veness, Marc G
Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, Stig
Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Ku-
maran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through
deep reinforcement learning. Nature, 518(7540):529–533, 2015. ISSN 0028-0836. URL
http://dx.doi.org/10.1038/nature14236.
Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap,
Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous Methods for Deep
Reinforcement Learning. In ICML, 2016.
Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized Experience
Replay. In ICLR, 2016.
John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust
Region Policy Optimization. In ICML, 2015.
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal
Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347, 2017.
David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur
Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. A general
reinforcement learning algorithm that masters chess, shogi, and go through self-play.
Science, 362(6419):1140–1144, 2018.
Adam Stooke and Pieter Abbeel. rlpyt: A Research Code Base for Deep Reinforcement
Learning in PyTorch. arXiv preprint arxiv:1909.01500, 2019.
Seiya Tokui, Ryosuke Okuta, Takuya Akiba, Yusuke Niitani, Toru Ogawa, Shunta Saito,
Shuji Suzuki, Kota Uenishi, Brian Vogel, and Hiroyuki Yamazaki Vincent. Chainer: A
Deep Learning Framework for Accelerating the Research Cycle. In KDD, 2019.
Hado Van Hasselt, Arthur Guez, and David Silver. Deep Reinforcement Learning with
Double Q-learning. In AAAI, 2016.
Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas.
Dueling Network Architectures for Deep Reinforcement Learning. In ICML, 2016.
Ziyu Wang, Victor Bapst, Nicolas Heess, Volodymyr Mnih, Remi Munos, Koray Kavukcuoglu,
and Nando de Freitas. Sample Efficient Actor-Critic with Experience Replay. In ICLR,
2017.
RJ Williams. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforce-
ment Learning. Machine Learning, 8(3-4):229–256, 1992.
Yuhuai Wu, Elman Mansimov, Shun Liao, Alec Radford, and John Schulman. OpenAI
Baselines: ACKTR & A2C. https://openai.com/blog/baselines-acktr-a2c/, 2017.
7Fujita, Nagarajan, Kataoka, and Ishikawa
Appendix A. Agents
ChainerRL implements several kinds of agents, supporting discrete-action agents, continuous-
action agents, recurrent agents, batch agents, and asynchronous agents. Asynchronous train-
ing, where an agent interacts with multiple environments asynchronously with a single set of
model parameters, is supported for A3C, ACER (Wang et al., 2017), N-step Q-learning, and
Path Consistency Learning (PCL). To train an asynchronous agent, one can simply initialize
an asynchronous agent and train it using experiments.train_agent_async. Batch training
refers to synchronous parallel training, where a single agent interacts with multiple environ-
ments synchronously in parallel, and is supported for all algorithms for which asynchronous
training is not supported. In ChainerRL, users can easily perform batch training of agents by
initializing an agent and using experiments.train_agent_batch_with_evaluation. Many
algorithms require additional infrastructure to support recurrent training, e.g., by storing
and managing the recurrent state, and managing sequences of observations as opposed to
individual observations. ChainerRL abstracts these difficulties away from the user, making it
simple to employ recurrent architectures for the majority of algorithms. Note that most of the
algorithms implemented in ChainerRL do not have support for recurrence or batch training
in their original published form. In ChainerRL, we have added this additional support for
most algorithms, as summarized in Table 1.
Algorithm Discrete Action Continuous Action Recurrent Model Batch Training CPU Async Training
DQN (Double DQN, SARSA, etc.) 3 3(NAF) 3 3 7
Categorical DQN 3 7 3 3 7
Rainbow 3 7 3 3 7
IQN (and Double IQN) 3 7 3 3 7
A3C 3 3 3 3(A2C) 3
ACER 3 3 3 7 3
NSQ (N-step Q-learning) 3 3(NAF) 3 7 3
PCL (Path Consistency Learning) 3 3 3 7 3
DDPG 7 3 3 3 7
PPO 3 3 3 3 7
TRPO 3 3 3 3 7
TD3 7 3 7 3 7
SAC 7 3 7 3 7
Table 1: Summarized list of ChainerRL algorithms and their additional supported features.
Appendix B. Reproducibility Results
For each of our reproducibility scripts, we provide the training times of the script (in our
repository), full tables of our achieved scores, and comparisons of these scores against those
reported in the literature. Though ChainerRL has high-quality implementations of dozens
of algorithms, we currently have created such “reproducibility scripts” for 9 algorithms. In
the Atari benchmark (Bellemare et al., 2013), we have successfully reproduced DQN, IQN,
Rainbow, and A3C. For the OpenAI Gym Mujoco benchmark tasks, we have successfully
reproduced DDPG, TRPO, PPO, TD3, and SAC.
The reproducibility scripts emphasize correctly reproducing evaluation protocols, which
are particularly relevant when evaluating Atari agents. Unfortunately, evaluation protocols
tend to vary across papers, and consequently results are often inconsistently reported across
8ChainerRL: A Deep Reinforcement Learning Library
the literature (Machado et al., 2018), significantly impacting results. The critical details of
standard Atari evaluation protocols are as follows:
Evaluation frequency The frequency (in timesteps) at which the evaluation phase occurs.
Evaluation phase length The number of timesteps in the offline evaluation.
Evaluation episode length The maximum duration of an evaluation episode.
Evaluation policy The policy to follow during an evaluation episode.
Reporting protocol Each intermediate evaluation phase outputs some score, representing
the mean score of all evaluation episodes during that evaluation phase. Papers typically
report scores according to one of the following reporting protocols:
1. best-eval : Papers using the best-eval protocol report the highest mean score across
all intermediate evaluation phases.
2. re-eval : Papers using the re-eval protocol report the score of a re-evaluation of
the network parameters that produced the best-eval.
During a typical Atari agent’s 50 million timesteps of training, it is evaluated periodically
in an offline evaluation phase for a specified number of timesteps before resuming training.
Since most papers report final results using the best model as determined by these periodic
evaluation phases, the frequency of evaluation is key, as it provides the author of a paper with
more models to select from when reporting final results. The length of the evaluation phase
is important, because shorter evaluation phases have higher variance in performance and
longer evaluation phases have less variance in performance. Again, since these intermediate
evaluations are used in some way when reporting final performance, the length of the
evaluation phase is important when reproducing results. The length of the evaluation
episodes can impact performance, as permitting the agent to have longer episodes may allow
it to accrue more points. Oftentimes, since the agent performs some form of exploratory
policy during training, the agent sometimes changes policies specifically for evaluations. Each
of the listed details, especially the reporting protocols, can significantly influence the results,
and thus are critical details to hold consistent for a fair comparison between algorithms.
Table 2 lists the results obtained by ChainerRL’s reproducibility scripts for DQN, IQN,
Rainbow, and A3C on the Atari benchmark, with comparisons against a published result.
Table 3 depicts the evaluation protocol used for each algorithm, with a citation of the source
paper whose results we compare against. Note that the results for the A3C (Mnih et al.,
2016) algorithm do not come from the original A3C paper, but from another (Fortunato
et al., 2018). For continuous-action algorithms, the results on OpenAI Gym MuJoCo tasks for
DDPG (Lillicrap et al., 2016), TRPO (Schulman et al., 2015), PPO (Schulman et al., 2017),
TD3 (Fujimoto et al., 2018), and SAC (Haarnoja et al., 2018) are reported in Table 4. For
all algorithms and environments listed in tables 2 and 4, we have released models trained
through our reproducibility scripts, which researchers can use.
The reproducibility scripts are produced through a combination of reading released
source code and studying published hyperparameters, implementation details, and evaluation
protocols. We also have extensive email correspondences with authors to clarify ambiguities,
omitted details, or inconsistencies that may exist in papers.
9Fujita, Nagarajan, Kataoka, and Ishikawa
As seen in both the Atari and MuJoCo reproducibility results, sometimes a reproduction
effort cannot be directly compared against the original paper’s reported results. For example,
the reported scores in the original paper introducing the A3C algorithm (Mnih et al., 2016)
utilize demonstrations that are not publicly available, making it impossible to accurately
compare a re-implementation’s scores to the original paper. In such scenarios, we seek out
high-quality published research (Fortunato et al., 2018; Henderson et al., 2018; Fujimoto
et al., 2018) from which faithful reproductions are indeed possible, and compare against
these.
10ChainerRL: A Deep Reinforcement Learning Library
DQN IQN Rainbow A3C
Game CRL Published CRL Published CRL Published CRL Published
Air Raid 6450.5 ± 5.9e+2 - 9933.5 ± 4.9e+2 - 6754.3 ± 2.4e+2 - 3923.8 ± 1.5e+2 -
Alien 1713.1 ± 2.3e+2 3069 12049.2 ± 8.9e+2 7022 11255.4 ± 1.6e+3 9491.7 2005.4 ± 4.3e+2 2027
Amidar 986.7 ± 1.0e+2 739.5 2602.9 ± 3.9e+2 2946 3302.3 ± 7.2e+2 5131.2 869.7 ± 7.7e+1 904
Assault 3317.2 ± 7.3e+2 3359 24315.8 ± 9.6e+2 29091 17040.6 ± 2.0e+3 14198.5 6832.6 ± 2.e+3 2879
Asterix 5936.7 ± 7.3e+2 6012 484527.4 ± 7.4e+4 342016 440208.0 ± 9.e+4 428200.3 9363.0 ± 2.8e+3 6822
Asteroids 1584.5 ± 1.6e+2 1629 3806.2 ± 1.5e+2 2898 3274.9 ± 8.4e+2 2712.8 2775.6 ± 3.3e+2 2544
Atlantis 96456.0 ± 6.5e+3 85641 937491.7 ± 1.6e+4 978200 895215.8 ± 1.3e+4 826659.5 836040.0 ± 4.7e+4 422700
Bank Heist 645.0 ± 4.7e+1 429.7 1333.2 ± 2.3e+1 1416 1655.1 ± 1.0e+2 1358.0 1321.6 ± 6.6e+0 1296
Battle Zone 5313.3 ± 2.9e+3 26300 67834.0 ± 5.1e+3 42244 87015.0 ± 1.3e+4 62010.0 7998.0 ± 2.6e+3 16411
Beam Rider 7042.9 ± 5.2e+2 6846 40077.2 ± 4.1e+3 42776 26672.1 ± 8.3e+3 16850.2 9044.4 ± 4.7e+2 9214
Berzerk 707.2 ± 1.7e+2 - 92830.5 ± 1.6e+5 1053 17043.4 ± 1.2e+4 2545.6 1166.8 ± 3.8e+2 1022
Bowling 52.3 ± 1.2e+1 42.4 85.8 ± 6.1e+0 86.5 55.7 ± 1.5e+1 30.0 31.3 ± 2.4e-1 37
Boxing 89.6 ± 3.1e+0 71.8 99.9 ± 2.1e-2 99.8 99.8 ± 1.3e-1 99.6 96.0 ± 1.9e+0 91
Breakout 364.9 ± 3.4e+1 401.2 665.2 ± 1.1e+1 734 353.0 ± 1.1e+1 417.5 569.9 ± 1.9e+1 496
Carnival 5222.0 ± 2.9e+2 - 5478.7 ± 4.6e+2 - 4762.8 ± 6.6e+2 - 4643.3 ± 1.2e+3 -
Centipede 5112.6 ± 6.9e+2 8309 10576.6 ± 1.7e+3 11561 8220.1 ± 4.6e+2 8167.3 5352.4 ± 3.3e+2 5350
Chopper Command 6170.0 ± 1.6e+3 6687 39400.9 ± 7.4e+3 16836 103942.2 ± 1.7e+5 16654.0 6997.1 ± 4.5e+3 5285
Crazy Climber 108472.7 ± 1.5e+3 114103 178080.2 ± 3.0e+3 179082 174438.8 ± 1.8e+4 168788.5 121146.1 ± 2.6e+3 134783
Demon Attack 9044.3 ± 1.8e+3 9711 135497.1 ± 1.5e+3 128580 101076.9 ± 1.1e+4 111185.2 111339.2 ± 6.3e+3 37085
Double Dunk -9.7 ± 1.8e+0 -18.1 5.6 ± 1.4e+1 5.6 -1.0 ± 7.9e-1 -0.3 1.5 ± 3.5e-1 3
Enduro 298.2 ± 5.4e+0 301.8 2363.6 ± 3.3e+0 2359 2278.6 ± 4.1e+0 2125.9 0.0 ± 0.e+0 0
Fishing Derby 11.6 ± 7.6e+0 -0.8 38.8 ± 4.3e+0 33.8 44.6 ± 5.1e+0 31.3 38.7 ± 1.6e+0 -7
Freeway 8.1 ± 1.3e+1 30.3 34.0 ± 0.e+0 34.0 33.6 ± 4.6e-1 34.0 0.0 ± 7.3e-3 0
Frostbite 1093.9 ± 5.5e+2 328.3 8196.1 ± 1.5e+3 4342 10071.6 ± 8.6e+2 9590.5 288.2 ± 2.9e+1 288
Gopher 8370.0 ± 1.1e+3 8520 117115.0 ± 2.8e+3 118365 82497.8 ± 5.6e+3 70354.6 9251.0 ± 1.8e+3 7992
Gravitar 445.7 ± 5.e+1 306.7 1006.7 ± 2.5e+1 911 1605.6 ± 1.9e+2 1419.3 244.5 ± 4.4e+0 379
Hero 20538.7 ± 2.0e+3 19950 28429.4 ± 2.4e+3 28386 27830.8 ± 1.3e+4 55887.4 36599.2 ± 3.5e+2 30791
Ice Hockey -2.4 ± 4.3e-1 -1.6 0.1 ± 2.0e+0 0.2 5.7 ± 5.4e-1 1.1 -4.5 ± 1.9e-1 -2
Jamesbond 851.7 ± 2.3e+2 576.7 26033.6 ± 3.8e+3 35108 24997.6 ± 5.6e+3 - 376.9 ± 2.6e+1 509
Journey Escape -1894.0 ± 5.8e+2 - -632.9 ± 9.7e+1 - -429.2 ± 4.4e+2 - -989.2 ± 4.2e+1 -
Kangaroo 8831.3 ± 6.8e+2 6740 15876.3 ± 6.4e+2 15487 11038.8 ± 5.8e+3 14637.5 252.0 ± 1.2e+2 1166
Krull 6215.0 ± 2.3e+3 3805 9741.8 ± 1.2e+2 10707 8237.9 ± 2.2e+2 8741.5 8949.3 ± 8.5e+2 9422
Kung Fu Master 27616.7 ± 1.3e+3 23270 87648.3 ± 1.1e+4 73512 33628.2 ± 9.5e+3 52181.0 39676.3 ± 2.4e+3 37422
Montezuma Revenge 0.0 ± 0.e+0 0.0 0.4 ± 6.8e-1 0.0 16.2 ± 2.2e+1 384.0 2.8 ± 6.3e-1 14
Ms Pacman 2526.6 ± 1.e+2 2311 5559.7 ± 4.5e+2 6349 5780.6 ± 4.6e+2 5380.4 2552.9 ± 1.9e+2 2436
Name This Game 7046.5 ± 2.0e+2 7257 23037.2 ± 2.e+2 22682 14236.4 ± 8.5e+2 13136.0 8646.0 ± 3.e+3 7168
Phoenix 7054.4 ± 1.9e+3 - 125757.5 ± 3.6e+4 56599 84659.6 ± 1.4e+5 108528.6 38428.3 ± 3.1e+3 9476
Pitfall -28.3 ± 2.1e+1 - 0.0 ± 0.e+0 0.0 -3.2 ± 2.9e+0 0.0 -4.4 ± 2.9e+0 0
Pong 20.1 ± 4.0e-1 18.9 21.0 ± 0.e+0 21.0 21.0 ± 6.4e-2 20.9 20.7 ± 3.9e-1 7
Pooyan 3118.7 ± 3.5e+2 - 27222.4 ± 9.9e+3 - 7772.7 ± 3.6e+2 - 4237.9 ± 5.8e+1 -
Private Eye 1538.3 ± 1.3e+3 1788 259.9 ± 1.0e+2 200 99.3 ± 5.8e-1 4234.0 449.0 ± 1.6e+2 3781
Qbert 10516.0 ± 2.6e+3 10596 25156.8 ± 5.3e+2 25750 41819.6 ± 1.9e+3 33817.5 18889.2 ± 7.6e+2 18586
Riverraid 7784.1 ± 6.8e+2 8316 21159.7 ± 8.0e+2 17765 26574.2 ± 1.8e+3 - 12683.5 ± 5.3e+2 -
Road Runner 37092.0 ± 3.e+3 18257 65571.3 ± 5.6e+3 57900 65579.3 ± 6.1e+3 62041.0 40660.6 ± 2.1e+3 45315
Robotank 47.4 ± 3.6e+0 51.6 77.0 ± 1.3e+0 62.5 75.6 ± 2.1e+0 61.4 3.1 ± 5.1e-2 6
Seaquest 6075.7 ± 2.3e+2 5286 26042.3 ± 3.9e+3 30140 3708.5 ± 1.7e+3 15898.9 1785.6 ± 4.1e+0 1744
Skiing -13030.2 ± 1.2e+3 - -9333.6 ± 7.4e+1 -9289 -10270.9 ± 8.6e+2 -12957.8 -13094.2 ± 3.7e+3 -12972
Solaris 1565.1 ± 6.e+2 - 7641.6 ± 8.2e+2 8007 8113.0 ± 1.2e+3 3560.3 3784.2 ± 3.5e+2 12380
Space Invaders 1583.2 ± 1.5e+2 1976 36952.7 ± 2.9e+4 28888 17902.6 ± 1.3e+4 18789.0 1568.9 ± 3.7e+2 1034
Star Gunner 56685.3 ± 1.0e+3 57997 182105.3 ± 1.9e+4 74677 188384.2 ± 2.3e+4 127029.0 60348.7 ± 2.6e+3 49156
Tennis -5.4 ± 7.6e+0 -2.5 23.7 ± 1.7e-1 23.6 -0.0 ± 2.4e-2 0.0 -12.2 ± 4.3e+0 -6
Time Pilot 5738.7 ± 9.0e+2 5947 13173.7 ± 7.4e+2 12236 24385.2 ± 3.5e+3 12926.0 4506.6 ± 2.8e+2 10294
Tutankham 141.9 ± 5.1e+1 186.7 342.1 ± 8.2e+0 293 243.2 ± 2.9e+1 241.0 296.7 ± 1.8e+1 213
Up N Down 11821.5 ± 1.1e+3 8456 73997.8 ± 1.7e+4 88148 291785.9 ± 7.3e+3 - 95014.6 ± 5.1e+4 89067
Venture 656.7 ± 5.5e+2 380.0 656.2 ± 6.4e+2 1318 1462.3 ± 3.4e+1 5.5 0.0 ± 0.e+0 0
Video Pinball 9194.5 ± 6.3e+3 42684 664174.2 ± 1.1e+4 698045 477238.7 ± 2.6e+4 533936.5 377939.3 ± 1.8e+5 229402
Wizard Of Wor 1957.3 ± 2.7e+2 3393 23369.5 ± 5.4e+3 31190 20695.0 ± 9.e+2 17862.5 2518.7 ± 5.1e+2 8953
Yars Revenge 4397.3 ± 2.1e+3 - 30510.0 ± 2.3e+2 28379 86609.9 ± 1.e+4 102557.0 19663.9 ± 6.6e+3 21596
Zaxxon 5698.7 ± 1.0e+3 4977 16668.5 ± 3.4e+3 21772 24107.5 ± 2.4e+3 22209.5 78.9 ± 6.8e+0 16544
# Higher scores 22 26 28 23 34 17 27 24
# Ties 1 4 1 3
# Seeds 5 1 3 1 3 1 5 3
Table 2: The performance of ChainerRL (± standard deviation) against published results
on Atari benchmarks.
11Fujita, Nagarajan, Kataoka, and Ishikawa
DQN IQN Rainbow A3C
Eval Frequency (timesteps) 250K 250K 250K 250K
Eval Phase (timesteps) 125K 125K 125K 125K
Eval Episode Length (time) 5 min 30 min 30 min 30 min
Eval Episode Policy = 0.05 = 0.001 = 0.0 N/A
Reporting Protocol re-eval best-eval re-eval best-eval
Table 3: Evaluation protocols used for the Atari reproductions. The evaluation protocols
of DQN, IQN, Rainbow, and A3C match the evaluation protocols used by Mnih
et al. (2015), Dabney et al. (2018), Hessel et al. (2017), and Fortunato et al. (2018),
respectively. An evaluation episode policy with an indicates that the agent
performs an -greedy evaluation.
DDPG (Fujimoto et al., 2018) TD3 (Fujimoto et al., 2018)
Environment CRL Published CRL Published
HalfCheetah-v2 10325.45 8577.29 10248.51 ± 1063.48 9636.95 ± 859.065
Hopper-v2 3565.60 1860.02 3662.85 ± 144.98 3564.07 ± 114.74
Walker2d-v2 3594.26 3098.11 4978.32 ± 517.44 4682.82 ± 539.64
Ant-v2 774.46 888.77 4626.25 ± 1020.70 4372.44 ± 1000.33
Reacher-v2 -2.92 -4.01 -2.55 ± 0.19 -3.60 ± 0.56
InvertedPendulum-v2 902.25 1000.00 1000.00 ± 0.0 1000.00 ± 0.0
InvertedDoublePendulum-v2 7495.56 8369.95 8435.33 ± 2771.39 9337.47 ± 14.96
TRPO (Henderson et al., 2018) PPO (Henderson et al., 2018) SAC (Haarnoja et al., 2018)
Environment CRL Published CRL Published CRL Published
HalfCheetah-v2 1474 ± 112 205 ± 256 2404 ± 185 2201 ± 323 14850.54 ~15000
Hopper-v2 3056 ± 44 2828 ± 70 2719 ± 67 2790 ± 62 2911.89 ~3300
Walker2d-v2 3073 ± 59 - 2994 ± 113 - 5282.61 ~5600
Ant-v2 - - - - 5925.63 ~5800
Swimmer-v2 200 ± 25 - 111 ± 4 - - -
Humanoid-v2 - - - - 7772.08 ~8000
Table 4: The performance of ChainerRL against published baselines on OpenAI Gym MuJoCo
benchmarks. For DDPG and TD3, each ChainerRL score represents the maximum
evaluation score during 1M-step training, averaged over 10 trials with different
random seeds, where each evaluation phase of ten episodes is run after every 5000
steps. For PPO and TRPO, each ChainerRL score represents the final evaluation of
100 episodes after 2M-step training, averaged over 10 trials with different random
seeds. For SAC, each ChainerRL score reports the final evaluation of 10 episodes
after training for 1M (Hopper-v2), 3M (HalfCheetah-v2, Walker2d-v2, and Ant-v2),
or 10M (Humanoid-v2) steps, averaged over 10 trials with different random seeds.
Since the original paper (Haarnoja et al., 2018) provides learning curves only, the
published scores are approximated visually from the learning curve. The sources of
the published scores are cited with each algorithm. We use the v2 environments,
whereas some published papers evaluate on the now-deprecated v1 environments.
12ChainerRL: A Deep Reinforcement Learning Library
Figure 2: The ChainerRL Visualizer. With the ChainerRL Visualizer, users can closely
investigate an agent’s behaviors within a browser window. top: Visualization of
a trained A3C agent on Breakout. bottom: Visualization of a C51 (Bellemare
et al., 2017) agent trained on Seaquest.
Appendix C. Visualizer Images
Figure 2 depicts some of the key features of the ChainerRL Visualizer for an actor-critic
algorithm and a distributional value-based algorithm. The top of the figure depicts a trained
A3C agent in the Atari game Breakout. With the visualizer, one can visualize the portions
of the pixel input that the agent is attending to as a saliency map (Greydanus et al., 2018).
Additionally, users can perform careful, controlled investigations of agents by manually
stepping through an episode, or can alternatively view rollouts of agents. Since A3C is an
actor-critic agent with a value function and a policy outputting a distribution over actions,
we can view the probabilities with which the agent will perform a specific action, as well as
the agent’s predicted state values. If the agent learns Q-values or a distribution of Q-values,
13Fujita, Nagarajan, Kataoka, and Ishikawa
the predicted Q-value or Q-value distribution for each action can be displayed, as shown in
the bottom of Figure 2.
Appendix D. Pseudocode
The set of algorithms that can be developed by combining the agent building blocks of
ChainerRL is large. One notable example is Rainbow (Hessel et al., 2017), which combines
double updating (Van Hasselt et al., 2016), prioritized replay (Schaul et al., 2016), N -step
learning, dueling architectures (Wang et al., 2016), and Categorical DQN (Bellemare et al.,
2017) into a single agent. The following pseudocode depicts the simplicity of creating and
training a Rainbow agent with ChainerRL.
1 import chainerrl as crl
2 import gym
3
4 q_func = crl . q_functions . D i s tr i b ut i o na l D ue l i ng D Q N (...) # dueling
5 crl . links . to_factorized_noisy ( q_func ) # noisy networks
6 # Prioritized Experience Replay Buffer with a 3 - step reward
7 per = crl . replay_buffers . P r io ri ti z ed Re p la yB uf f er ( num_step_return =3 ,...)
8 # Create a rainbow agent
9 rainbow = crl . agents . CategoricalDoubleDQN ( per , q_func ,...)
10 num_envs = 5 # Train in five environments
11 env = crl . envs . Multip roces sVecto rEnv (
12 [ gym . make ( " Breakout " ) for _ in range ( num_envs ) ])
13
14 # Train the agent and collect evaluation statistics
15 crl . experiments . t r a i n _ a g e n t _ b a t c h _ w i t h _ e v a l u a t i o n ( rainbow , env , steps =...)
We first create a distributional dueling Q-function, and then in a single line, convert it
to a noisy network. We then initialize a prioritized replay buffer configured to use N -step
rewards. We pass this replay buffer to ChainerRL’s built-in CategoricalDoubleDQN agent
to produce a Rainbow agent. Moreover, with ChainerRL, users can easily specify the number
of environments in which to train the Rainbow agent in synchronous parallel processes, and
the experiments module will automatically manage the training loops, evaluation statistics,
logging, and saving of the agent.
14You can also read
